641 Views
SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Pim van Ooij, PhD
Post-doc Researcher
Amsterdam University Medical Center, AMC
Martijn Froeling, PhD
Post-doc researcher
University Medical Center Utrecht
Emile Farag, MD
MD
Amsterdam University Medical Center
R. Nils Planken, MD, PhD
Radiologist
Amsterdam University Medical Center
Tim Leiner, MD, PhD
Professor of Radiology
Utrecht University Medical Center
Background:
Wall shear stress (WSS) is a clinically relevant biomarker for degradation of the aortic wall in congenital heart disease [1]. The quantification of WSS from 4D flow MRI demands a precise segmentation of the aortic lumen for definition of the aortic wall [2]. Manual segmentation is time-consuming, laborious and inhibits clinical use of WSS. We hypothesize that WSS estimated with aortic segmentations created by a machine learning approach is comparable with WSS estimated using manual segmentations.
Methods:
4D flow MRI datasets (spatiotemporal resolution: 2.5x2.5x2.5 mm3/±42 ms, TE/TR/FA: 2.1ms/3.4ms/8°, k-t PCA acceleration R=8) were acquired on a 3T MRI Ingenia system (Philips, Best, the Netherlands) in 25 healthy volunteers and 42 patients with bicuspid valves (of which 22 patients with repaired coarctation). The aortic lumen was segmented semi-automatically (Mimics, Materialise, Leuven) from images created by multiplying the absolute velocity with the magnitude followed by averaging over all time frames using thresholding, watershed and manual drawing/erasing algorithms [3]. The magnitude, velocity and magnitude x velocity images were averaged over time and used, in addition to the manually created masks, as input for the convolutional neural network (CNN). The control and patient datasets were combined and evenly split up in n=47 (70%), n=13 (20%) and n=7 (10%) for training, validation and testing of the CNN. A hybrid 3D-UNet/ResNet CNN (figure 1) was used for training with 32 features in the first layer and 1024 in the deepest (38 convoluting layers with 3.8 million parameters in total). Cross-entropy and soft-dice were used as loss function. Training of the CNN was performed using 500 epochs using an ADAM optimizer (Nvidia TitanXP). Peak systolic WSS was calculated using wall definition by 1) the manual segmentations (WSSM) and 2) the machine learning segmentations (WSSML). For voxel-by-voxel Bland-Altman and orthogonal regression analysis, WSSML was rigidly registered and interpolated to WSSM [3].
Results:
Training of the CNN took 2.1 hours for 500 epochs but the loss and error curves show that training converged after ±100 epochs (figure 2). In figure 3 three test exemplary datasets are displayed. The patient/control Dice-similarity for the test, validation and training data was 0.892/0.901, 0.896/0.913 and 0.941/0.919. The mean difference, limits of agreement, Pearson’s r, slope and intercept between WSSM and WSSML for the test/all datasets was 0.01±0.06/0.01±0.05Pa, 0.59±0.24/0.63±0.24Pa, 0.79±0.08/0.76±0.09, 0.99±0.15/0.97±0.12 and 0.02±0.11/0.03±0.09Pa.
Conclusion:
WSS estimated with automatically created segmentations using a machine learning approach are comparable with WSS estimated with manually created segmentations. The clinical application of 4D flow MRI-derived WSS is greatly improved by the ability to automatically create aortic segmentations.